Audit 2/3: the instrument - datasheet-sourced simulator and tooling

[enoch85]

The instrument, split from the product: the simulator and dev tooling. Stacked on the production PR; the validation tests that read both land in the next PR.

• The plant is datasheet-sourced. COP from an exergy-efficiency fit on each machine's own EN 14511 points (lands in OpenEnergyMonitor's measured 45–55%-of-Carnot band); capacity from the published envelope; houses sized from each pump's declared Pdesignh; every physical constant SOURCED (a document you can open) or ASSUMED (with measured sensitivity).
• The elpatron fires where NIBE arms it (each pump's factory start-addition), not at EffektGuard's −1500 floor. Re-measured, DM asymptotes at the hardware value — exactly what Swedish forum reports describe — and the F-124 cold-snap headline shrinks to its honest size (1.4–1.7× forced aux, house held in band).
• The harness runs production code, not lookalikes: the real BillingPeriodAccumulator, the real GE-Spot adapter on committed real SE4 prices, the real decision layers. Break production billing and --dst fails (the fall-back day must bill 25 hours).
• The tariff figure is what Ellevio bills: hourly means, one peak per day, night half-weighting.
• Scenario modes and their current verdicts: nominal/--selftest/--dst PASS all five houses; --coldsnap and --undersized FAIL on the aux-over-physics bound — the instrument still detects the saturation defect, it just no longer exaggerates it.
• Tooling: the hardcoded-values ratchet scanner, the week-long live-HA watcher (fixed sensor id, env-driven login), dead demo scripts removed.